Transition Network Analysis

The same sequence data can be represented as different types of networks. TNA constructs a standard first-order Markov model with row-normalized transition probabilities. FTNA retains raw transition counts without normalization — suited for descriptive analysis when sequence lengths vary (online tutorial). CTNA builds an undirected network from state co-occurrence within sequences. ATNA applies exponential decay based on the gap between states, with forward, backward, and bidirectional variants (vignette). WTNA adjusts for sequence length by inversely weighting transitions. Each model encodes different assumptions. The choice is substantive, not computational.

The methodological core of TNA is validation at four levels — model, edges, nodes, and networks — described in the TNA primer and formalized in the R package paper.

Model reliability comes first. The reliability() function performs split-half validation: the sequence data is randomly divided into two halves, a TNA model is estimated from each, and the two networks are compared over many iterations (default 1,000). Four metrics quantify agreement: Pearson correlation between edge weights, mean and median absolute difference, and Bray-Curtis dissimilarity. On the group_regulation dataset, split-half Pearson correlations average 0.993 — the model structure is highly reproducible. If reliability is low, no downstream analysis can rescue the findings.

Edge validation uses bootstrap pruning or disparity filtering. Bootstrap pruning resamples sequences with replacement, estimates a model from each sample, and retains edges that appear consistently — yielding a p-value and confidence interval per edge. The disparity filter retains edges whose weights are significant relative to a null model of uniform distribution.

Node validation addresses centrality stability. TNA computes nine centrality measures: InStrength, OutStrength, Closeness (In/Out/combined), Betweenness, random-walk Betweenness, Diffusion, and Clustering coefficient. Case-dropping stability analysis removes increasing proportions of sequences and tracks correlation with original centralities. If correlations hold above 0.7 when half the data is removed, the estimates are reliable.

Network validation uses permutation testing for group comparison. Sequences from both groups are pooled, group labels shuffled, and a test statistic computed for each permutation to obtain a p-value and effect size at the edge or network level. The group analysis tutorial and comparison guide demonstrate this with worked examples.

The four layers form a chain where each guards against a different source of error. Every claim — this transition is significant, this state is central, these groups differ — has been tested against a null hypothesis. The analysis is replicable because validation is built into the method, not left to the analyst’s judgment.

Community detection partitions the network into subsets of states with denser internal connections (vignette). Clique detection identifies maximally connected subgraphs. Sequence-level clustering groups individuals by edit distance, Hamming distance, or longest common subsequence into behaviorally distinct subgroups (clustering tutorial; book chapter). Network-level metrics — density, reciprocity, mean distance, centralization, degree distribution — provide a quantitative fingerprint for comparison across conditions or time periods.